454 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 2

454 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 2, FEBRUARY 2011

Exploiting Sparse User Activity inMultiuser Detection

Hao Zhu, Student Member, IEEE, and Georgios B. Giannakis, Fellow, IEEE

Abstract—The number of active users in code-division multipleaccess (CDMA) systems is often much lower than the spreadinggain. The present paper exploits fruitfully this a priori infor-mation to improve performance of multiuser detectors. A low-activity factor manifests itself in a sparse symbol vector withentries drawn from a finite alphabet that is augmented by thezero symbol to capture user inactivity. The non-equiprobablesymbols of the augmented alphabet motivate a sparsity-exploitingmaximum a posteriori probability (S-MAP) criterion, which isshown to yield a cost comprising the ℓ2 least-squares errorpenalized by the 𝑝-th norm of the wanted symbol vector(𝑝 = 0, 1, 2). Related optimization problems appear in variableselection (shrinkage) schemes developed for linear regression,as well as in the emerging field of compressive sampling (CS).The contribution of this work to such sparse CDMA systemsis a gamut of sparsity-exploiting multiuser detectors tradingoff performance for complexity requirements. From the vantagepoint of CS and the least-absolute shrinkage selection operator(Lasso) spectrum of applications, the contribution amounts tosparsity-exploiting algorithms when the entries of the wantedsignal vector adhere to finite-alphabet constraints.

Index Terms—Sparsity, multiuser detection, compressive sam-pling, Lasso, sphere decoding.

I. INTRODUCTION

MULTIUSER detection (MUD) algorithms play a majorrole for mitigating multi-access interference present

in code-division multiple access (CDMA) systems; see e.g.,[17] and references therein. These well-appreciated MUDalgorithms simultaneously detect the transmitted symbols ofall active user terminals. However, they require knowledgeof which terminals are active, and exploit no possible user(in)activity.

In this paper, MUD algorithms are developed when the ac-tive terminals are unknown and the activity factor (probabilityof each user being active) is low – a typical scenario in tacticalor commercial CDMA systems deployed. The inactivity per

Paper approved by X. Wang, the Editor for Multiuser Detection andEqualization of the IEEE Communications Society. Manuscript receivedSeptember 21, 2009; revised April 2, 2010, July 3, 2010, and August 30,2010.

This work was supported by the NSF grants CCF-0830480, 1016605,and ECCS-0824007, 1002180; and also through the collaborative partic-ipation in the Communications and Networks Consortium sponsored bythe U. S. Army Research Laboratory under the Collaborative TechnologyAlliance Program, Cooperative Agreement DAAD19-01-2-0011. The U. S.Government is authorized to reproduce and distribute reprints for Governmentpurposes notwithstanding any copyright notation thereon. Part of this workwas presented at the IEEE Intl. Symp. on Info. Theory, Seoul, South Korea,June 2009.

The authors are with the Department of Electrical and Computer Engineer-ing, University of Minnesota, 200 Union Street SE, Minneapolis, MN 55455,USA (e-mail: {zhuh, georgios}@umn.edu).

Digital Object Identifier 10.1109/TCOMM.2011.121410.090570

user can be naturally incorporated by augmenting the under-lying alphabet with an extra zero constellation point. Lowactivity thus implies a sparse symbol vector to be detected.With non-equiprobable symbols in the augmented alphabet,the optimal sparsity-embracing MUD naturally suggests amaximum a posteriori (MAP) criterion for detection. Sparsityhas been used for estimating parameters of communicationsystems, see e.g., [1], [3], [6], [10], but not for multiuserdetection1.

Reconstruction of sparse signals has become very popularrecently, especially after the emergence of the compressivesampling (CS) theory; see e.g., [2], [4], [7], [16] and refer-ences therein. Exploiting sparsity has also received growinginterest in regression problems because it offers parsimoniousstatistical models that prevent data overfitting. In particular,the least-absolute shrinkage and selection operator (Lasso) hasbeen applied for variable selection (VS) to diverse applicationsinvolving sparse signals [15]. Lasso regression amounts toregularizing the ordinary least-squares (LS) with the ℓ1 normof the wanted vector, thus retaining the attractive features ofboth VS and LS regression. However, neither CS nor Lassohave dealt with sparse signals under finite-alphabet constraints;and this is the subject of the present work.

Different from sparse signal reconstruction or regression,the multiuser detectors developed in this paper must accountfor sparsity but also for the finite alphabet of the desiredsymbol vector. Taking into consideration the sparsity in MUDunder integer constraints however, incurs complexity that isexponential in the vector length. To cope with this challenge,we will exploit sparsity to either relax or judiciously searchover subsets of the alphabet. The resultant MUD algorithmstrade off optimality in detection error performance with com-putational complexity. In a nutshell, this paper’s contributionis the development of efficient algorithms for MUD undersparsity and finite-alphabet constraints.

Also in the CS literature a number of works is available todeal with the performance of recovering sparse sequences ofsize exceeding the number of observations; see e.g., [4], [5]and references therein. These works establish the minimumnumber of observations needed to estimate sparse vectorswith overwhelming probability. In the design of practicalCDMA systems however, one is also interested in savingbandwidth and power resources. This becomes possible byreducing the size of the required spreading gain, which in turnreduces latency and energy consumption. Taking advantage ofthe low activity factor, enables designs with spreading gains

1Recently, along with the conference pre-cursor of this paper in [20], workrelated to sparse sphere decoding was reported independently in [14].

0090-6778/11$25.00 c⃝ 2011 IEEE

ZHU and GIANNAKIS: EXPLOITING SPARSE USER ACTIVITY IN MULTIUSER DETECTION 455

smaller than the number of candidate users. Efficient multiuserdetectors are also developed here for this kind of under-determined systems encountered in CDMA communications.

The rest of the paper is organized as follows. SectionII presents the model and formulates the problem; whileSection III introduces the optimal sparsity-exploiting MAPdetector. (Sub-) optimal algorithms are developed in SectionsIV and V. Section VI lists several extensions. All the detectionalgorithms are tested and compared numerically in SectionVII, before concluding in Section VIII.

II. MODELING AND PROBLEM STATEMENT

Consider the uplink of a CDMA system with 𝐾 userterminals and spreading gain 𝑁 . Assume first that 𝑁 ≥ 𝐾 .The under-determined case (𝑁 < 𝐾) will also be addressedlater on in Section VI-B. Suppose the system has a relativelylow activity factor, which analytically means that each termi-nal is active with probability (w.p.) 𝑝𝑎 < 1/2 per symbol,and the events “active” are independent across symbols andacross users. The case of correlated (in)activity of users acrosssymbols will be dealt with in Section VI-A. Let 𝑏𝑘 ∈ 𝒜denote the symbol, drawn from a finite alphabet by the 𝑘-thuser, when active; otherwise, 𝑏𝑘 = 0. Incorporating possible(in)activity per user is equivalent to having 𝑏𝑘 take values froman augmented alphabet 𝒜𝑎 := 𝒜∪{0}.

The access point (AP) receives the superimposed modulated(quasi-) synchronous signature waveforms through (possiblyfrequency-selective) fading channels in the presence of ad-ditive white Gaussian noise (AWGN); and projects on theorthonormal space spanned by the aggregate waveforms toobtain the received chip samples collected in the 𝑁 × 1vector y. With the 𝐾 × 1 vector b containing the symbolsof all (active and inactive) users, the canonical input-outputrelationship is, see e.g., [17, Sec. 2.9]

y = Hb + w (1)

where H is an 𝑁×𝐾 matrix capturing transmit-receive filters,spreading sequences, channel impulse responses, and timingoffsets; and the 𝑁 × 1 vector w is the AWGN. Without lossof generality (w.l.o.g.), w can be scaled to have unit variance.Note that (1) holds for quasi-synchronous systems too, whererelative user asynchronism is bounded to a few chips persymbol, provided that: either i) user transmissions includeguard bands to eliminate inter-symbol interference (ISI); orii) the received vector y collects only the chips of each userthat belong to the part of the common symbol interval underconsideration.

The low activity factor implies that b is a sparse vector.However, the AP is neither aware of the positions nor thenumber of zero entries in b. In order to perform multiuserdetection (MUD) needed to determine the optimal b, the APmust account for the augmented alphabet 𝒜𝑎, i.e., consider allthe possible candidate vectors b ∈ 𝒜𝐾

𝑎 . This way, the MUDalso determines the 𝑘-th user’s activity captured by the extraconstellation point 𝑏𝑘 = 0. Supposing that the AP has thechannel matrix H available (e.g., via training), the goal ofthis paper is to detect the optimal b given the received vectory by exploiting the sparsity of active users.

Fig. 1. Wireless sensors access a UAV.

To motivate this sparsity-exploiting MUD setup in theCDMA context, consider a set of terminals wishing to linkwith a common AP. Suppose that the AP acquires the fullmatrix H (with all terminals active) during a training phase.Those channels may include either non-dispersive or multipathfading, and are assumed invariant during the coherence time,which is typically larger compared to the symbol period. Eachterminal accesses the channel randomly, and the AP receivesthe superposition of signals from the active terminals only.The AP is interested in determining both the active terminalsand the symbols transmitted.

Another scenario where H is known and sparsity-exploitingMUD is well motivated, entails an unmanned aerial vehicle(UAV) collecting information from a ground wireless sensornetwork (WSN) placed over a grid, as depicted in Fig. 1.As the UAV flies over the grid of sensors, it collects thesignals from a random subset of them. If the channel fadingis predominantly affected by path loss, the UAV can acquireH based on the relative AP-to-sensor distances. Again, theUAV faces the problem of determining both identities of activesensors, and the symbols each active sensor transmits.

With this problem setup in mind, we will develop differentMUD strategies, which account for the low activity factor.First, we will look at the maximum a posteriori probability(MAP) optimum MUD that exploits the sparsity present.

III. SPARSITY-EXPLOITING MAP DETECTOR

The goal is to detect b in (1), given a prescribed activityfactor, the received vector y, and the matrix H. Recall thoughthat the low activity factor leads to a sparse b, i.e., each entry𝑏𝑘 is more likely to take the value 0 from the alphabet. Becauseentries {𝑏𝑘}𝐾𝑘=1 are non-equiprobable, the optimal detector inthe sense of minimizing the detection error rate is the MAPone.

Aiming at a sparsity-aware MAP criterion, consider firstthe prior probability for b. For simplicity in exposition,suppose for now that each terminal transmits binary symbolswhen active, i.e., 𝒜 = {±1}. (It will become clear lateron that all sparsity-cognizant MUD schemes are applicableto finite alphabets of general constellations, not necessarilybinary.) If 𝑏𝑘 takes values from {−1, 0, 1}, with correspondingprobabilities {𝑝𝑎/2, 1 − 𝑝𝑎, 𝑝𝑎/2}, and since each entry 𝑏𝑘 isindependent from 𝑏𝑘′ for 𝑘 ∕= 𝑘′, the prior probability for b


can be expressed as

Pr(b) =

𝐾∏𝑘=1

Pr(𝑏𝑘) = (1 − 𝑝𝑎)𝐾−∥b∥0(𝑝𝑎/2)∥b∥0 (2)

where ∥b∥0 denotes the ℓ0 (pseudo) norm that is by definitionequal to the number of non-zero entries in the vector b. Upontaking logarithms, (2) yields

ln Pr(b) = −𝜆∥b∥0 + 𝐾 ln(1 − 𝑝𝑎) (3)

where

𝜆 := ln1 − 𝑝𝑎𝑝𝑎/2

. (4)

Since low activity factor means 𝑝𝑎 < 1/2, it follows readilyfrom (4) that 𝜆 > 0. With the prior distribution of b in (3),the sparsity-aware MAP (S-MAP) detector is

bMAP =arg maxb∈𝒜𝐾

𝑎

Pr(b∣y)=arg minb∈𝒜𝐾

𝑎

−ln 𝑝(y∣b)−ln Pr(b)

= arg minb∈𝒜𝐾

𝑎

1

2∥y −Hb∥22 + 𝜆∥b∥0 (5)

where the last equality follows from (3) and the Gaussianityof w. Hence, the S-MAP detection task amounts to findingthe vector in the constraint set 𝒜𝐾

𝑎 , which minimizes the costin (5).

Two remarks are now in order.Remark 1: (General constellations). Beyond binary alpha-bets, it is easy to see why the S-MAP detector in (5) appliesto more general constellations, including pulse amplitudemodulation (PAM), phase-shift keying (PSK), and quadratureamplitude modulation (QAM). Specifically, for general 𝑀 -aryconstellations with 𝑀 ≥ 2 it suffices to adjust accordinglythe prior probability as a function of ∥b∥0 in (2). Thiswill render the S-MAP MUD in (5) applicable to general𝑀 -ary constellations, provided that 𝜆 in (4) is replaced by𝜆 := ln 1−𝑝𝑎

𝑝𝑎/𝑀.

Remark 2: (Scale 𝜆 as a function of 𝑝𝑎). The definition in (4)reveals the explicit relationship of 𝜆 with the activity factor𝑝𝑎. Different from CS and VS approaches, where 𝜆 is a tuningparameter often chosen with cross-validation techniques asa function of the data size 𝑁 and 𝐾 , here it is directlycoupled with the user activity factor 𝑝𝑎. Such a couplingcarries over even when users have distinct activity factors.Specifically, if the user 𝑘 is active w.p. 𝑝𝑎,𝑘, then the term𝜆∥b∥0 := 𝜆

∑𝐾𝑘=1 ∣𝑏𝑘∣ in (5) should change to

∑𝐾𝑘=1 𝜆𝑘∣𝑏𝑘∣,

where 𝜆𝑘 is defined as in (4) with 𝑝𝑎,𝑘 substituting 𝑝𝑎. Thisuser-specific regularization will be used in Section VI-A toadaptively estimate user activity factors on-the-fly, and thusenable sparsity-aware MUD which accounts for correlated user(in)activity across the time slots.

With the variable 𝑏𝑘 only taking values from {±1, 0}, itholds for 𝑝 ≥ 1 that

∥b∥0 =

𝐾∑𝑘=1

∣𝑏𝑘∣𝑝 = ∥b∥𝑝𝑝, ∀ b ∈ 𝒜𝐾𝑎 . (6)

Hence, the S-MAP detector (5) for binary transmissions isequivalent to

bMAP = arg minb∈𝒜𝐾

𝑎

1

2∥y −Hb∥22 + 𝜆∥b∥𝑝𝑝 , ∀ 𝑝 ≥ 1. (7)

Notice that the equivalence between (5) and (7) is based onthe norm equivalence in (6), which holds only for constantmodulus constellations. Although the cost in (7) will turn outto be of interest on its own, it is not an S-MAP detector fornon-constant modulus constellations.

Interestingly, since low-activity factor implies a positive𝜆, the problem (7) entails a convex cost function, comparedto the non-convex one in (5). In lieu of the finite-alphabetconstraint, the criterion of (7) consists of the least-squares(LS) cost regularized by the ℓ𝑝 norm, which in the context oflinear regression has been adopted to mitigate data overfitting.In contrast, the LS estimator only considers the goodness-of-fit, and thus tends to overfit the data. Shrinking the LSestimator, by penalizing its size through the ℓ𝑝 norm, typicallyoutperforms LS in practice. For example, the Lasso adoptsthe ℓ1 norm through which it effects sparsity [15]. In theMUD context for CDMA systems with low-activity factor,the vector b has a sparse structure, which motivates well thisregularizing strategy. What is distinct and interesting here isthat this penalty-augmented LS approach under finite-alphabetconstraints emerges naturally as the logarithm of the prior inthe S-MAP detector.

However, the finite-alphabet constraint renders the solutionof (7) combinatorially complex. For general H and y, thesolution of (7) requires exhaustive search over all the 3𝐾

feasible points, with the complexity growing exponentiallyin the problem dimension 𝐾 . Likewise, for general 𝑀 -aryalphabets the complexity incurred by (5) is 𝒪((𝑀 +1)𝐾). Onthe other hand, many (sub-) optimal alternatives are availablein the MUD literature; see e.g., [17, Ch. 5-7]. Similarly here,we will develop different (sub-) optimal algorithms to tradeoff complexity for probability of error performance in sparsity-exploiting MUD alternatives.

Since the exponential complexity of MUD stems from thefinite-alphabet constraint, one reduced-complexity approachis to solve the unconstrained convex problem, and thenquantize the resultant soft decision to the nearest point inthe alphabet. This approach includes the sub-optimal linearMUD algorithms (decorrelating and minimum mean-squareerror (MMSE) detectors). Another approach is to search over(possibly a subset of) the alphabet lattice directly as inthe decision-directed detectors or the sphere decoders [18].Likewise, it is possible to devise (sub-) optimal algorithmsfor solving the S-MAP MUD problem (7) along these twocategories. First, we will present the sparsity-exploiting MUDalgorithms by relaxing the finite-alphabet constraint.

IV. RELAXED S-MAP DETECTORS

In addition to offering an S-MAP detector for constantmodulus constellations, the cost in (7) is convex. Thus, byrelaxing the combinatorial constraint, the optimization prob-lem (7) can be solved efficiently by capitalizing on convexity.As mentioned earlier, this problem is similar to the penalty-augmented LS criterion that is used for VS in linear regression,where the choice of 𝑝 is important for controlling the shrinkingeffect, that is the degree of sparsity in the solution. Next, wewill develop detectors for two choices of 𝑝, and compare themin terms of complexity and performance.


A. Linear Ridge MUD

The choice 𝑝 = 2 is a popular one in statistics, well-known as Ridge regression. Its popularity is mainly due tothe fact that it can regularize the LS solution while retainingits closed-form expression as a linear function of the datay. A relaxed detection algorithm for S-MAP MUD can bedeveloped accordingly with 𝑝 = 2, what we term Ridgedetector (RD). Ignoring the finite-alphabet constraint, theoptimal solution of (7) for 𝑝 = 2 takes a linear form

bRD = arg minb

1

2∥y −Hb∥22 + 𝜆∥b∥22

= (H𝑇H + 2𝜆I)−1H𝑇y. (8)

In addition to its simplicity, and different from LS, theexistence of the inverse in (8) is ensured even for ill-posedor under-determined problems; i.e., when H is rank deficientor fat (𝑁 < 𝐾). Notice that bRD takes a form similar tothe linear MMSE multiuser detector, with the parameter 𝜆replacing the noise variance, and connecting the activity factorwith the degree of regularization applied to the matrix H𝑇H.

Upon quantizing each entry of the soft decision bRD withthe operator

𝒬𝜃(𝑥) := sign(𝑥)11(∣𝑥∣ ≥ 𝜃) (9)

where 𝜃 > 0, sign(𝑥) = 1(−1) with 𝑥 > (<)0, and 11denoting the indicator function, the hard RD is

bRD = 𝒬𝜃(bRD) = 𝒬𝜃

((H𝑇H + 2𝜆I)−1H𝑇y

). (10)

Because the detector in (8) is linear, it is possible toexpress linearly its soft output bRD with respect to (w.r.t.) theinput symbol vector b. Based on this relationship, one cansubsequently derive the symbol error rate (SER) of the harddetected symbols in bRD as a function of the quantizationthreshold 𝜃. These steps will be followed next to obtain theperformance of the RD.

1) Performance Analysis: Letting b denote the vectortransmitted, and substituting y = Hb + w into (8) yields

bRD = (H𝑇H + 2𝜆I)−1H𝑇 (Hb + w) = Gb + w′ (11)

where G := I − 2𝜆(H𝑇H + 2𝜆I)−1, and the colorednoise w′ := (H𝑇H + 2𝜆I)−1H𝑇w is zero-mean Gaussianwith covariance matrix Σ𝑤′ := 𝐸{w′(w′)𝑇 } = (H𝑇H +2𝜆I)−2H𝑇H.

It follows readily from (11) that the 𝑘-th entry of bRD

satisfies

𝑏RD𝑘 = 𝐺𝑘𝑘 ��𝑘 +

∑ℓ ∕=𝑘

𝐺𝑘ℓ��ℓ + 𝑤′𝑘 (12)

where 𝐺𝑘ℓ and 𝑤′𝑘 are the (𝑘, ℓ)-th and 𝑘-th entries of G

and w′, respectively. The last two terms in the right-hand sideof (12) capture the multiuser interference-plus-noise effect,which has variance

𝜎2𝑘 = var

⎧⎨⎩∑ℓ ∕=𝑘

𝐺𝑘ℓ��ℓ + 𝑤′𝑘

⎫⎬⎭ =

∑ℓ ∕=𝑘

𝐺2𝑘ℓ𝑝𝑎 + Σ𝑤′,𝑘𝑘 (13)

where Σ𝑤′,𝑘𝑘 denotes the (𝑘, 𝑘)-th entry of Σ𝑤′ .With the interference-plus-noise term being approximately

Gaussian distributed, deciphering ��𝑘 from (12) amounts to

detecting a ternary deterministic signal in the presence of zero-mean, Gaussian noise of known variance. Hence, the symbolerror rate (SER) for the 𝑘-th entry using the quantization rulein (10) entry-wise can be analytically obtained as

𝑃 RD𝑒,𝑘 = 2(1 − 𝑝𝑎)𝑄

(𝜃

𝜎𝑘

)+ 𝑝𝑎𝑄

(𝐺𝑘𝑘 − 𝜃

𝜎𝑘

)(14)

where 𝑄(𝜇) := (1/√

2𝜋)∫∞𝜇

exp(−𝜈2/2)𝑑𝜈 denotes theGaussian tail function.

The SER in (14) is a convex function of the threshold 𝜃.Thus, taking the first-order derivative w.r.t. 𝜃 and setting itequal to zero yields the optimal threshold for the 𝑘-th entryas

𝜃𝑘 =𝐺𝑘𝑘

2+

𝜎2𝑘

𝐺𝑘𝑘𝜆 . (15)

The corresponding minimum SER becomes [cf. (14) and (15)]

𝑃 RD𝑒,𝑘 =2(1 − 𝑝𝑎)𝑄

(𝐺𝑘𝑘

2𝜎𝑘+

𝜆𝜎𝑘

𝐺𝑘𝑘

)+𝑝𝑎𝑄

(𝐺𝑘𝑘

2𝜎𝑘− 𝜆𝜎𝑘

𝐺𝑘𝑘

).

(16)

As the CDMA system signal-to-noise ratio (SNR) goes toinfinity, asymptotically we have G → I and Σ𝑤′ → 0, sothe optimal threshold in (15) approaches 0.5. The numericaltests in Section VII will also confirm that selecting 𝜃𝑘 = 0.5approaches the minimum SER 𝑃 RD

𝑒,𝑘 over the range of SNRvalues encountered in most practical settings.

The clear advantage of RD-MUD is its simplicity as alinear detector. However, using the ℓ2 norm for regularization,the RD-MUD inherently introduces a Gaussian prior for theunconstrained symbol vector and is thus not effecting sparsityin bRD; see also [15]. This renders the performance of RDdependent on the quantization threshold 𝜃 – a fact also corrob-orated by the simulations in Section VII. These considerationsmotivate the ensuing development of an alternative relaxed S-MAP algorithm, which accounts for the sparsity present inb.

B. Lasso-based MUD

Another popular regression method is the Lasso one, whichregularizes the LS cost with the ℓ1 norm. In the Bayesianformulation, regularization with the ℓ1 norm corresponds toadopting a Laplacian prior for b [15]. The nice feature ofLasso-based regression is that it ensures sparsity in the resul-tant estimates. The degree of sparsity depends on the valueof 𝜆, which is selected here using the a priori informationavailable on the activity factor [cf. (4)]. The optimal solutionof (7) for 𝑝 = 1 without the finite-alphabet constraint yieldsthe Lasso detector (LD) as

bLD = arg minb

1

2∥y −Hb∥22 + 𝜆∥b∥1. (17)

While a closed-form solution is impossible for general H,the minimization in (17) is a quadratic programming (QP)problem that can be readily accomplished using available QPsolvers, such as SeDuMi [13]. Upon slicing the solution in(17), we obtain the detection result as

bLD = 𝒬𝜃(bLD). (18)


The larger 𝜆 is, the more sparse bLD becomes [cf. (4)]. Thisis intuitively reasonable, because 𝜆 is inversely proportionalto the activity factor 𝑝𝑎. Since the Lasso approach (17)yields sparse estimates systematically, and can be obtainedvia QP solvers in polynomial time, LD is a competitive MUDalternative. Lack of a closed-form solution prevents analyticalevaluation of the SER, which will be tested using simulationsin Section IV.

Remark 3: So far, we assumed 𝒜 = {±1} to ensureequivalence of the ℓ𝑝-norm regularized S-MAP detector (7)with the more general one in (5). However, the sub-optimalalgorithms of this section ignore the finite-alphabet constraint,and just rely on the convexity of the cost function in (7) tooffer MUD schemes that can be implemented efficiently, eitherin linear closed-form or through quadratic programming. Infact, starting from (7) and forgoing its equivalence with (5),the RD and LD relaxations of (7) apply also for general 𝑀 -ary alphabets for any 𝑀 > 2. Of course, the quantizationthresholds required for slicing the soft symbol estimates inorder to obtain hard symbol estimates must be modified inaccordance with the corresponding constellation. For non-constant modulus transmissions, the cost in (7) favors low-energy (close to the origin) constellation points, but this effectis mitigated by the judicious selection of the quantizationthresholds.

Note that forgoing the equivalence of (7) with (5) is lessof an issue for RD and LD because the major limitationof these simple relaxation-based algorithms is their sub-optimality, which emerges because they do not account forthe finite-alphabet symbol constraints. Next, MUD algorithmsare developed to minimize the S-MAP cost while adhering tothe constraint in (5) explicitly.

V. S-MAP DETECTORS WITH LATTICE SEARCH

User symbols in this section are drawn from an 𝑀 -ary PAM alphabet 𝒜 = {±1,±3, . . . ,±(𝑀 − 1)}, with𝑀 even. Consider also reformulating the S-MAP problemin (5) using the QR decomposition of the matrix H (as-sumed here to be square or tall with full column rank) asH = QR, where R is a 𝐾 × 𝐾 upper triangular matrix,and Q is an 𝑁 × 𝐾 unitary matrix. Substituting this QRdecomposition into (5), and left multiplying with the uni-tary Q inside the LS cost, the S-MAP detector becomes:bMAP = arg minb∈𝒜𝑘

𝑎

12

∥∥Q𝑇y −Q𝑇 (QR)b∥∥22

+ 𝜆∥b∥0 =arg minb∈𝒜𝐾

𝑎

12∥y′ −Rb∥22 + 𝜆∥b∥0, where y′ := Q𝑇y; or,

after using the definitions of the norms,


𝑎

𝐾∑𝑘=1

⎧⎨⎩1

2

(𝑦′𝑘 −

𝐾∑ℓ=𝑘

𝑅𝑘ℓ𝑏ℓ

)2

+ 𝜆∣𝑏𝑘∣0

⎫⎬⎭ .

(19)

Although the optimal solution of (19) still incurs expo-nential complexity, the upper triangular form of R enablesdecomposition of (19) into sub-problems involving only scalarvariables. As it will be seen later in Section V-A, the S-MAPproblem accepts a neat closed-form solution in the scalar case(𝐾 = 1). This is instrumental for the development of efficient(near-) optimal algorithms searching over the finite-alphabet

induced lattice. One such sub-optimal MUD algorithm isdescribed next.

A. Sparsity-Exploiting Decision-Directed MUD

Close look at (19) reveals that once the estimates {��ℓ}𝐾ℓ=𝑘+1

are available, the optimal ��𝑘 can be obtained by minimizingthe cost corresponding to the 𝑘-th summand of (19). This leadsto the per-symbol optimal decision-directed detector (DDD),which following related schemes in different contexts, couldbe also called successive interference cancellation or decision-feedback decoding, see e.g., [11, Sec. 9.4] and [17, Ch. 7]. Themain difference here is that b is sparse.

The DDD algorithm relies on back substitution to decom-pose the overall S-MAP cost into 𝐾 sub-costs each dependenton a single scalar variable and accepting a closed-form solu-tion. Specifically, supposing the symbols {��DDD

ℓ }𝐾ℓ=𝑘+1 havebeen already detected, the DDD algorithm detects the 𝑘-thsymbol as

��DDD𝑘 = arg min

𝑏𝑘∈𝒜𝑎

1

2

(𝑦′𝑘 −

𝐾∑ℓ=𝑘+1

𝑅𝑘ℓ��DDDℓ − 𝑅𝑘𝑘𝑏𝑘

)2

+ 𝜆∣𝑏𝑘∣0 . (20)

This minimization problem entails only one scalar variabletaking one of (𝑀 + 1) possible points in 𝒜𝑎. Thus, theminimum is found after comparing the costs correspondingto these (𝑀 + 1) values. Appendix A proves that this optimalsolution can be found in closed form as

��DDD𝑘 = ⌊𝑏LS

𝑘 ⌉11 (2𝑏LS𝑘 ⌊𝑏LS

𝑘 ⌉ − ⌊𝑏LS𝑘 ⌉2 − 2𝜆/𝑅2

𝑘𝑘 > 0)

(21)

where 𝑏LS𝑘 := (𝑦′

𝑘−∑𝐾

ℓ=𝑘+1 𝑅𝑘ℓ��DDDℓ )/𝑅𝑘𝑘 , and ⌊⋅⌉ quantizes

to the nearest point in 𝒜. The simple implementation steps aretabulated as Algorithm 1.

Algorithm 1 (DDD): Input y′, R, 𝜆, and 𝑀 even.Output bDDD.

1: for 𝑘 = 𝐾,𝐾 − 1, . . . , 1 do2: (Back substitution) Compute the unconstrained LS solution

𝑏LS𝑘 := (𝑦′

𝑘 −∑𝐾ℓ=𝑘+1𝑅𝑘ℓ��

DDDℓ )/𝑅𝑘𝑘 .

3: (Quantize to 𝒜) Set ��DDD𝑘 := ⌊𝑏LS

𝑘 ⌉.4: (Compare with 0) If 2𝑏LS

𝑘 ⌊𝑏LS𝑘 ⌉ − ⌊𝑏LS

𝑘 ⌉2 − 2𝜆/𝑅2𝑘𝑘 ≤ 0, then

set ��DDD𝑘 := 0.

5: end for

When R is diagonal, Algorithm 1 yields the optimal S-MAPdetection result; i.e., bMAP = bDDD for this case. However,since the DDD detects symbols sequentially, it is prone toerror propagation, especially at low SNR values. The errorpropagation can be mitigated by preprocessing and orderingmethods [17, Ch. 7]. Also similar to all related detectors thatrely on back substitution, error performance of the sparseDDD will be analyzed assuming there is no error propagation.For the special case of 𝑀 = 2, Appendix B shows that underthis assumption, the SER for (21) becomes

𝑃DDD𝑒,𝑘 = 2(1 − 𝑝𝑎)𝑄

( ∣𝑅𝑘𝑘∣2

+𝜆

∣𝑅𝑘𝑘∣)

+𝑝𝑎𝑄

( ∣𝑅𝑘𝑘∣2

− 𝜆

∣𝑅𝑘𝑘∣)

. (22)


For a general 𝑀 -ary constellation, it is also possible toapproximate the SER using the union bound.

Because it accounts for the finite-alphabet constraint, sparseDDD outperforms the relaxed detectors of the previous section– a fact that will be confirmed also by simulations. How-ever, the error propagation emerging at medium-low SNRdegrades the sparse DDD performance when compared tothe optimal but computationally complex S-MAP detector. Asa compromise, a branch-and-bound type of MUD algorithmis developed next, to attain (near-) optimal performance byexploiting the finite-alphabet and sparsity constraints, at theprice of increased complexity compared to DDD.

B. Sparsity-Exploiting Sphere Decoding-based MUD

Sphere decoding (SD) algorithms have been widely used formaximum-likelihood (ML) demodulation of multiuser and/ormultiple-input multiple-output (MIMO) transmissions. Givena PAM or QAM alphabet, SD yields (near-) ML performanceat polynomial average complexity; see e.g., [11, Sec. 5.2].However, different from the ML-optimal SD that minimizes anLS cost, the S-MAP problem (19) entails also a regularizationterm to account for sparsity in b. Although the resultantalgorithm will be termed sparse sphere decoder (SSD), itsearches in fact within an “ℓ0-norm regularized sphere,” whichis not a sphere but a hyper-solid.

The goal is to find the unknown 𝐾 × 1 vector b ∈ 𝒜𝐾𝑎 ,

which minimizes the distance metric [cf. (19)]

𝐷𝐾1 (b) :=

𝐾∑𝑘=1

⎧⎨⎩1

2

(𝑦′𝑘 −

𝐾∑ℓ=𝑘

𝑅𝑘ℓ𝑏ℓ

)2

+ 𝜆∣𝑏𝑘∣0

⎫⎬⎭ . (23)

For a large enough threshold 𝜏 , candidate vectors (and thusthe minimizer of 𝐷𝐾

1 too) satisfy2

𝐷𝐾1 (b) < 𝜏 (24)

that specifies a hyper-solid inside which the wanted minimizermust lie. Define now [cf. (21)]

𝜌𝑘 :=

(𝑦′𝑘 −

𝐾∑ℓ=𝑘+1

𝑅𝑘ℓ𝑏ℓ

)/𝑅𝑘𝑘, 𝑘 = 𝐾,𝐾 − 1 ⋅ ⋅ ⋅ 1

(25)

where 𝜌𝐾 := 𝑦′𝐾/𝑅𝐾𝐾 . Note that 𝜌𝑘 depends on {𝑏ℓ}𝐾ℓ=𝑘+1.

Using (25), the hyper-solid in (24) can be expressed as

𝐷𝐾1 (b) =

𝐾∑𝑘=1

{𝑅2

𝑘𝑘

2(𝜌𝑘 − 𝑏𝑘)

2+ 𝜆∣𝑏𝑘∣0

}< 𝜏 (26)

or, in a more compact form as 𝐷𝐾1 (b) =

∑𝐾𝑘=1 𝑑𝑘(𝑏𝑘) < 𝜏 ,

where 𝑑𝑘(𝑏𝑘) := (𝑅2𝑘𝑘/2) (𝜌𝑘 − 𝑏𝑘)

2+ 𝜆∣𝑏𝑘∣0. In addition to

the overall metric 𝐷𝐾1 assessing a candidate vector b ∈ 𝒜𝐾

𝑎 ,as well as the per entry metric 𝑑𝑘 for each candidate symbol𝑏𝑘 ∈ 𝒜𝑎, it will be useful to define the accumulated metric𝐷𝐾

𝑘 :=∑𝐾

ℓ=𝑘 𝑑ℓ(𝑏ℓ) corresponding to the 𝐾−𝑘+1 candidatesymbols from entry 𝐾 down to entry 𝑘.

2At initialization, 𝜏 is set equal to ∞ so that (24) is always satisfied.

Per entry 𝑘, which subsequently will be referred to as level𝑘, eq. (26) implies a set of inequalities:

Level 𝑘 : 𝑑𝑘(𝑏𝑘) <𝜏 − 𝐷𝐾𝑘+1 ,

for 𝑘 = 𝐾,𝐾 − 1 . . . 1, (27)

with 𝐷𝐾𝐾+1 := 0. SSD relies on the Schnorr-Euchner (SE)

enumeration, see e.g., [9], properly adapted here to account forthe ℓ0-norm regularization. SE capitalizes on the inequalities(27) to search efficiently over all possible vectors b withentries belonging to 𝒜𝑎. Any candidate b ∈ 𝒜𝐾

𝑎 obeying the𝐾 inequalities in (27) for a given 𝜏 , will be termed admissible.Threshold 𝜏 is reduced after each admissible b is identified, aswill be detailed soon. The SE-based SSD amounts to a depth-first tree search, which seeks and checks candidate vectorsstarting from entry (level) 𝐾 and working downwards to entry1 per candidate vector.

At level 𝐾 , SE search chooses 𝑏𝐾 = ⌊𝜌𝐾⌉11(2𝜌𝐾⌊𝜌𝐾⌉ −⌊𝜌𝐾⌉2 − 2𝜆/𝑅2

𝐾𝐾 > 0), which we know from (21) is theconstellation point yielding the smallest 𝑑𝐾 . If this choiceof 𝑏𝐾 does not satisfy the inequality (27) with 𝑘 = 𝐾 ,no other constellation point will satisfy it either, and theminimizer of 𝐷𝐾

1 in (23) must lie outside3 the hyper-solidpostulated by (24). If this choice of 𝑏𝐾 satisfies (27), SEproceeds to level 𝐾 − 1 in which (25) is used first with𝑘 = 𝐾 − 1 to find 𝜌𝐾−1 that depends on the chosen 𝑏𝐾from level 𝐾; subsequently, 𝑏𝐾−1 is selected as 𝑏𝐾−1 =⌊𝜌𝐾−1⌉11(2𝜌𝐾−1⌊𝜌𝐾−1⌉ − ⌊𝜌𝐾−1⌉2 − 2𝜆/𝑅2

𝐾−1,𝐾−1 > 0).If this choice of 𝑏𝐾−1 does not satisfy (27) with 𝑘 = 𝐾 − 1,then we move back to level 𝐾 , and select 𝑏𝐾 equal to theconstellation point yielding the second smallest 𝑑𝐾 , and soon; otherwise, we proceed to level 𝐾 − 2. Continuing thisprocedure down to level 1, yields the first candidate vector b,which is deemed admissible since it has entries belonging to𝒜𝑎 and also satisfying (24). This candidate is stored, and thethreshold is updated to 𝜏 := 𝐷𝐾

1 (b).Then, the search proceeds looking for a better candidate.

Now at level 1, we move up to level 2 and choose 𝑏2 equal tothe constellation point yielding the second smallest cost 𝑑2.If this 𝑏2 satisfies (27) at level 2 with the current 𝜏 , we movedown to level 1 to update the value of 𝑏1 (note that 𝑏2 hasjust been updated and {𝑏ℓ}𝐾ℓ=3 are equal to the correspondingentries in b). If (27) at level 2 is not satisfied with the current𝜏 , we move up to level 3 to update the value of 𝑏3, and soon.

Finally, when it fails to find any other admissible candidatesatisfying (27), the search stops, and the latest admissiblecandidate b is the optimal bMAP solution sought. With𝜏 = ∞, the first found admissible candidate b is the bDDD

solution of Section V-A.Before summarizing the SSD steps, it is prudent to elaborate

on the ordered enumeration of the constellation points perlevel, which in fact constitutes the main difference of SSDrelative to SD. In lieu of the 0 constellation point and theℓ0 norm, SE in SD enumerates the PAM symbols per level𝑘 in the order of increasing cost as: {𝑏𝑘, 𝑏𝑘 + 2Δ𝑘, 𝑏𝑘 −2Δ𝑘, 𝑏𝑘 + 4Δ𝑘, 𝑏𝑘 − 4Δ𝑘, ⋅ ⋅ ⋅ }

∩𝒜, with 𝑏𝑘 := ⌊𝜌𝑘⌉ andΔ𝑘 := sign(𝜌𝑘 − 𝑏𝑘). (With ⌊𝜌𝑘⌉ yielding the smallest 𝑑𝑘,

3This will never happen with 𝜏 = ∞ in (24).


if Δ𝑘 = 1, then ⌊𝜌𝑘⌉ + 2 yields the second smallest 𝑑𝑘 and⌊𝜌𝑘⌉−2 the third; and the other way around, if Δ𝑘 = −1.) SDeffects such an ordered enumeration by alternately updating𝑏𝑘 = 𝑏𝑘+2Δ𝑘 and Δ𝑘 = −Δ𝑘−sign(Δ𝑘) [9]. To demonstratehow SSD further accounts for the ℓ0-norm regularization andthe augmented alphabet of S-MAP which includes 0, let𝑏(𝑖)𝑘 ∈ 𝒜𝑎 denote the symbol for level 𝑘 incurring the 𝑖-th

smallest (𝑖 = 1, 2, . . . ,𝑀 +1) cost 𝑑𝑘. If 𝑏(𝑖)𝑘 ∈ 𝒜, then 𝑏

(𝑖+1)𝑘

will be either 0 or 𝑏(𝑖)𝑘 + 2Δ𝑘. If 𝑑𝑘(0) < 𝑑𝑘(𝑏

(𝑖)𝑘 + 2Δ𝑘),

then the next symbol in the ordered enumeration should be𝑏(𝑖+1)𝑘 = 0, and an auxiliary variable 𝑏

(𝑐)𝑘 is used to cache the

subsequent symbol in the order as 𝑏(𝑖+2)𝑘 = 𝑏

(𝑖)𝑘 + 2Δ𝑘. With

𝑏(𝑖+1)𝑘 = 0, the auxiliary variable allows the wanted 𝑏

(𝑖+2)𝑘 at

the next enumeration step to be retrieved from 𝑏(𝑐)𝑘 .

Similar to SD, the ordered enumeration pursued by SSD perlevel implies a corresponding order in considering all b ∈ 𝒜𝐾

𝑎 ,which leads to a repetition-free and exhaustive search of alladmissible candidate vectors. At the same time, the hyper-solid postulated by (24) shrinks as 𝜏 decreases, until no otheradmissible vector can be found. This guarantees that the SSDoutputs the vector with the smallest 𝐷𝐾

1 , and thus the optimalsolution to (19). The SSD algorithm can be summarized inthe following six steps 1–6 tabulated as Algorithm 2.

Remark 4: SSD inherits all the attractive features of SD [9].Specifically, during the search one basically needs to store𝐷𝐾

𝑘 per level 𝑘. Its in place update for each 𝑏𝑘 candidateimplies that SSD memory requirements are only linear in 𝐾 . Inaddition, the computational efficiency of SSD (relative to thatof ML which is 𝒪((𝑀 +1)𝐾)) stems from four factors: (i) theDDD solution provides an admissible initialization reducingthe search space at the outset; (ii) the recursive search enabledby the QR decomposition gives rise to the causally dependentinequalities (27), which restrict admissible candidate entries tochoices that even decrease over successive depth-first passesof the search; (iii) ordering per level increases the likelihoodof finding “near-optimal admissible” candidates early, whichmeans quick and sizeable shrinkage of the hyper-solid, andthus fast convergence to the S-MAP optimal solution; and(iv) metrics involved in the search can be efficiently reusedsince children of the same level in the tree share the alreadycomputed accumulated metric of the “partial path” from thislevel to the root.

Compared to other sub-optimal detection schemes proposedin previous sections, the SSD algorithm can return the S-MAP optimal solution possibly at exponential complexity,unless one stops the search at the affordable complexity –case in which the solution is only ensured to be near-optimal.Fortunately, at medium-high SNR, both SD and SSD return theoptimal solution at average complexity which is polynomial(typically cubic). Moreover, SSD can be generalized to providesymbol-by-symbol soft output with approximate a posterioriprobabilities, as is the case with the SD; see e.g., [11, Chapter5].

VI. GENERALIZATIONS OF S-MAP DETECTORS

Up to now, four sparsity-exploiting MUD algorithms havebeen developed to solve the integer program associated withthe linear model in (1). The present section will present

Algorithm 2 (SSD): Input 𝜏 = ∞, y′, R, 𝜆, and 𝑀 even.Output the solution bMAP := bSSD to (19).

1: (Initialization) Set 𝑘 := 𝐾, 𝐷𝐾𝑘+1 := 0.

2: Compute 𝜌𝑘 as in (25), set 𝑏𝑘 := ⌊𝜌𝑘⌉, 𝑏(𝑐)𝑘 := 0, and Δ𝑘 :=sign(𝜌𝑘 − 𝑏𝑘).If 2𝜌𝑘𝑏𝑘 − 𝑏2𝑘 − 2𝜆/𝑅2

𝑘𝑘 < 0, then// symbol 0 yields smaller 𝑑𝑘 than ⌊𝜌𝑘⌉

Set 𝑏(𝑐)𝑘 := 𝑏𝑘, and 𝑏𝑘 := 0.End if and go to Step 3.

3: If 𝑑𝑘(𝑏𝑘) := (𝑅2𝑘𝑘/2)(𝜌𝑘 − 𝑏𝑘)

2 + 𝜆∣𝑏𝑘∣0 ≥ 𝜏 −𝐷𝐾𝑘+1, then

go to Step 4. // outside hyper-solid in (24)Else if ∣𝑏𝑘∣ > 𝑀 − 1, then

go to Step 6. // inside hyper-solid in (24),but outside 𝒜𝑎

Else if 𝑘 > 1, thencompute 𝐷𝐾

𝑘 := 𝐷𝐾𝑘+1 + 𝑑𝑘(𝑏𝑘); set 𝑘 := 𝑘 − 1, and go

to Step 2. // go the next level(deeper in the tree)

Else go to Step 5. // 𝑘 = 1, at the tree’s bottomEnd if

4: If 𝑘 = 𝐾, then terminateElse set 𝑘 := 𝑘 + 1, go to Step 6.End if

5: (An admissible b is found)Set 𝜏 := 𝐷𝐾

2 + 𝑑1(𝑏1), bSSD := b, and 𝑘 := 𝑘 + 1; then go toStep 6.

6: (Enumeration at level 𝑘 proceeds to the candidate symbol nextin the order)If 𝑏𝑘 = 0, then

Retrieve the next (based on cost 𝑑𝑘 ordering) symbol 𝑏𝑘 :=𝑏(𝑐)𝑘 , and set 𝑏(𝑐)𝑘 := 𝐹𝐿𝐴𝐺.

Else set 𝑏𝑘 := 𝑏𝑘 + 2Δ𝑘 , and Δ𝑘 := −Δ𝑘 − sign(Δ𝑘).

If 𝑏(𝑐)𝑘 ∕= 𝐹𝐿𝐴𝐺 and 2𝜌𝑘𝑏𝑘 − 𝑏2𝑘 − 2𝜆/𝑅2𝑘𝑘 < 0, then

// 0 yields smaller 𝑑𝑘 than 𝑏𝑘Set 𝑏(𝑐)𝑘 := 𝑏𝑘, and 𝑏𝑘 := 0.

End ifEnd if and go to Step 3.

interesting generalizations to account for correlated user(in)activity across symbols, and under-determined CDMAsystems.

A. Exploiting user (in)activity across symbols

Sparsity-aware detectors for the linear model in (1) wereso far developed on a symbol-by-symbol basis, which doesnot account for the fact that user (in)activity typically persistsacross multiple symbols. To this end, user activity across timecan for instance be thought of as a Markov chain with twostates (active-inactive). Once a user terminal starts transmittingto the AP, it becomes more likely to stay active for thenext symbol slot too; and likewise, inactive once it stopstransmitting. In this model, the state transition probability fromeither one state to the other is relatively much smaller thanthat of staying unchanged, and this manifests itself to the saiddependence of user (in)activities across time.

Admittedly, MUD schemes accounting for this dependencemust process the aggregation of data vectors y obeying (1)across slots. With 𝑁𝑠 denoting the number of slots, the numberof unknowns (𝐾𝑁𝑠) can grow prohibitively with 𝑁𝑠. Oneapproach to cope with this “curse of dimensionality” is via


dynamic programming, which can take advantage of the factthat correlation is only present between two consecutive slots;see e.g., [17, Sec. 4.2]. However, for 𝑀 -ary alphabets, theresultant sequential detector requires evaluating per symbolslot the path weights of all (𝑀 +1)𝐾 possible symbol vectors.This high computational burden is impractical for real-timeimplementations.

The proposed alternative to bypass this challenge stemsfrom the observation that for a given slot 𝑡 the influence ofall previous slots {𝑡′}𝑡−1

𝑡′=0 on the S-MAP detection rule isreflected only in the prior probability of each user being activeat time 𝑡; i.e., user 𝑘’s current (and time-varying) activityfactor 𝑝𝑎,𝑘(𝑡). The natural means to capture this influenceonline is to track each user’s activity factor using the recursiveLS (RLS) estimator [12, Ch. 12], based on activity factorsfrom previous slots; that is,

𝑝𝑎,𝑘(𝑡)=arg min𝑝

𝑡−1∑𝑡′=0

𝛽𝑡,𝑡′𝑘 (𝑝 − ∣��𝑘(𝑡′)∣0)2, 𝑡 = 1, . . . (28)

where ��𝑘(𝑡′) denotes user 𝑘’s detected symbol at time 𝑡′, andthe so-called “forgetting-factor” 𝛽𝑡,𝑡′

𝑘 describes the effectivememory (data windowing). A popular choice is the exponen-tially decaying window, for which 𝛽𝑡,𝑡′

𝑘 := 𝛽𝑡−𝑡′𝑘 for some

0 ≪ 𝛽𝑘 < 1. Accordingly, (28) can expressed in closed form,recursively as

𝑝𝑎,𝑘(𝑡)=1 − 𝛽𝑘

𝛽𝑘

(𝑡−1∑𝑡′=0

𝛽𝑡−𝑡′𝑘 ∣��𝑘(𝑡′)∣0

)/(1 − 𝛽𝑡

𝑘)

=𝛽𝑘 − 𝛽𝑡

𝑘

1 − 𝛽𝑡𝑘

𝑝𝑎,𝑘(𝑡 − 1)+1 − 𝛽𝑘

1 − 𝛽𝑡𝑘

∣��𝑘(𝑡 − 1)∣0, 𝑡 = 1, . . .

(29)

where the last equality comes from back substitution of𝑝𝑎,𝑘(𝑡 − 1). The choice of 𝛽𝑘 critically depends on the(in)activity correlation between consecutive slots. In the ex-treme case where user (in)activities across slots are inde-pendent, the infinite-memory window (𝛽𝑘 = 1) is optimal,and (29) reduces to the simple online time-averaging estimate𝑝𝑎,𝑘(𝑡) = 1

𝑡

∑𝑡−1𝑡′=0 ∣��𝑘(𝑡′)∣0.

Adapting the user activity factors allows one to weighentries of ℓ0-norm regularization which in turn affects the priorprobability in the S-MAP detector (5) through the coefficient𝜆𝑘(𝑡) corresponding to 𝑝𝑎,𝑘(𝑡) (cf. Remark 2). Note that whenthe correlation across time is strong, it is possible that 𝑝𝑎,𝑘(𝑡)can approach 1, case in which 𝜆𝑘(𝑡) is not guaranteed tostay positive. This will cause problems to the relaxed S-MAP detectors of Section IV, as those schemes rely on theconvexity of the cost function in (7). However, the DDDand SSD algorithms will remain operational, because theyrely on enumeration per symbol (in DDD) or per group ofsymbols within a sphere (in SSD). Note also that with 𝜆 < 0the regularization term in the minimization of (19) is non-positive; hence, 𝜆 < 0 encourages searching over the non-zeroconstellation points (and thus discourages sparsity), whereas𝜆 > 0 promotes sparsity.

B. Under-determined CDMA Systems

Minimal-size spreading sequences, even smaller than thenumber of users, is well motivated for bandwidth and power

savings. Without finite-alphabet constraints on the wantedvector, results available in the CS literature guarantee recoveryof sparse signals from a few observations; see e.g., [4], [5] andreferences therein. Specifically, [4] shows that if the vectorof interest is sparse (or compressible) over a known basis,then it is possible to reconstruct it with very high accuracyfrom a small number of random linear projections at least inthe ideal noise-free case. For non-ideal observations corruptedwith unknown noise of bounded perturbation, [5] provides anupper bound on the reconstruction error, which is proportionalto the noise variance for a sufficiently sparse signal. However,CS theory pertains to sparse analog-valued signals. Moreover,the noise considered in a practical communication system istypically Gaussian, or generally drawn from a distributionhaving possibly unbounded support. Therefore, existing resultsfrom the CS literature do not carry over to the present context.

Nevertheless, it is still interesting to consider extensions ofall the (sub-)optimal sparsity-exploiting MUD algorithms toan under-determined CDMA system with 𝑁 < 𝐾 , where theobservation matrix H becomes fat. Consider first the two typesof relaxed S-MAP detectors. The RD-MUD in (10) clearlyworks when 𝑁 < 𝐾 , since the 2𝜆I term inside the inversionrenders the overall matrix full rank. However, since the RDis a linear detector, it is expected to lose identifiability inthe under-determined case, similar to the MMSE detectorsfor the sparsity-agnostic MUD schemes. Similarly, the LDproblem (17) is also solvable for a fat H matrix, as the Lassoproblem in CS. However, neither of them accounts for theaugmented finite-alphabet constraint present in the original S-MAP problem (7).

The S-MAP detectors with lattice search are challengingto implement when 𝑁 < 𝐾 . The main obstacle is the QRdecomposition of the fat matrix H, which yields the uppertriangular matrix R of the same dimension 𝑁 ×𝐾 . Instead ofa single unknown symbol, now the sparse DDD must optimizeover the last (𝑁 − 𝐾 + 1) symbols in b. However, apartfrom exhaustive search there is no low-complexity methodto solve the aforementioned problem involving (𝑁 − 𝐾 +1) variables, because sub-optimal alternatives introduce severeerror propagation.

The same problem appears also with the SSD. To tacklethe under-determined case, the generalized SD in [9] fixes thelast (𝑁 −𝐾) symbols of b and relies on the standard SSD todetect the remaining 𝐾 symbols that minimize a cost similarto the one in (23). Repeating this search for every choice of thelast (𝑁 −𝐾) symbols, yields eventually the overall optimumvector. The complexity of the latter is exponential in (𝑁−𝐾),regardless of the SNR. Recently, an alternative SD approach toavoid this exponential complexity has been developed for theunder-determined case [8]. This algorithm takes advantage ofthe fact that for constant-modulus constellations the usual LScost can be modified without affecting optimality, by addingthe ℓ2-norm b𝑇b constant for every vector in the alphabet.This extra term allows one to obtain an equivalent full-ranksystem on which the standard SD algorithm can be applied.This efficient method can be readily extended to handle non-constant modulus constellations.

Interestingly, for our S-MAP detectors in (7) with latticesearch of binary transmitted symbols, the norm term needed


for regularization comes naturally from the Bernoulli prior.Specifically, with 𝑝 = 2 the reformulated S-MAP detectors in(7) can be equivalently written as


𝑎

1

2[b𝑇 (H𝑇H + 2𝜆I)b− 2y𝑇Hb]

= arg minb∈𝒜𝐾

𝑎

1

2∥y′ − Rb∥22 (30)

where R is the full rank 𝐾 ×𝐾 upper triangular matrix suchthat R𝑇 R = H𝑇H + 2𝜆I, and y′ := R−𝑇H𝑇y. Utilizingthe metric of (30), the back substitution of DDD and thelattice point search of SSD can be implemented easily. Hence,these S-MAP detectors can be readily extended to under-determined systems. In this way, all the (sub-)optimal S-MAPdetectors are applicable with less observations than unknownsin a CDMA system with low activity factor, but their SERperformance will certainly be affected. In Section VII, we willprovide simulated performance comparisons of the differentoptimal and sub-optimal S-MAP algorithms proposed, for avariable number of observations.

C. Group Lassoing block activity

The last generalization considered pertains to user(in)activity in a (quasi-)synchronous block fashion, whereduring a block of 𝑁𝑠 symbol slots, user 𝑘 remains (in)activeindependently from other users and across blocks. Concatenatethe 𝐾 user symbols across 𝑁𝑠 time slots in the 𝐾×𝑁𝑠 matrixB := [b(1) . . .b(𝑁𝑠)], where b(𝑡) collects the symbols of all𝐾 users at slot 𝑡, and likewise for the receive-data matrixY as well as the noise matrix W, both of size 𝑁 × 𝑁𝑠.With these definitions, the counterpart of (1) for this blockmodel is Y = HB + W. Letting the 𝑁𝑠 × 1 vector b𝑘 :=[𝑏𝑘(1), . . . , 𝑏𝑘(𝑁𝑠)]

𝑇 collect the 𝑁𝑠 symbols of user 𝑘, it isuseful to consider it drawn from an augmented (due to possibleinactivity) block alphabet 𝒜𝑎,𝑁𝑠 := 𝒜𝑁𝑠

∪{0𝑁𝑠}. Assumingagain binary transmissions, the S-MAP block detector willnow yield

BMAP = arg minb𝑘∈𝒜𝑎,𝑁𝑠

1

2∥Y −HB∥2𝐹 +

𝐾∑𝑘=1

𝜆𝑏√𝑁𝑠

∥b𝑘∥0

= arg minb𝑘∈𝒜𝑎,𝑁𝑠

1

2∥Y −HB∥2𝐹 +

𝐾∑𝑘=1

𝜆𝑏∥b𝑘∥2 (31)

where 𝜆𝑏 := 1√𝑁𝑠

ln(

1−𝑝𝑎

𝑝𝑎/2𝑁𝑠

).

Similar to (7), the convex reformulation of the cost in (31)will lead to what is referred to in statistics as Group Lasso[19], which effects group sparsity on a block of symbols(b𝑘 in our case). This Group-Lasso based formulation isparticularly handy for under-determined CDMA systems. Infact, the unconstrained version of (31) can be solved first tounveil the nonzero rows (i.e., the support) of B, with improvedreliability as 𝑁𝑠 increases. Subsequently, standard sparsity-agnostic MUD schemes can be run on the estimated set ofactive users. Note that such a two-step approach works in theunder-determined case, and also reduces number of symbolsto be detected per time slot.

5 10 15 20 25

10−3

10−2

Average SNR(dB)

SE

R

P

e

Optimal θk

θ = 0.5θ = 0.35θ = 0.65

Fig. 2. SER vs. average SNR (in dB) for RD-MUD with 𝑁 = 32 and𝐾 = 20 and different quantization threshold 𝜃’s.

VII. SIMULATIONS

We simulate a CDMA system with 𝐾 = 20 users, eachwith activity factor 𝑝𝑎 = 0.3. Random sequences with length𝑁 = 32 are used for spreading. We consider non-dispersiveindependent Rayleigh fading channels between AP and users,where the channel gain 𝑔𝑘 of the 𝑘-th user is Rayleighdistributed with variance 𝐸[𝑔2𝑘] = 𝜎2. Thus, the averagesystem SNR is set to be 𝜎2 since the AWGN w has unitvariance.

Test Case 1 (Quantization thresholds for RD). First, we testthe RD scheme with different quantization thresholds 𝜃 in(9). The optimum threshold for the 𝑘-th symbol is obtainedas in (15) per channel realization H. The resulting SER iscompared for four choices of 𝜃: 0.5, 0.35, 0.65, and 𝜃𝑘. Thetheoretical minimum SER 𝑃 RD

𝑒,𝑘 in (16) using the optimum 𝜃 isalso added for comparison. Fig. 2 shows that the SER curvewith 𝜃 = 0.5 comes very close to the one of the optimal𝜃𝑘, and thus constitutes a near-optimal choice in practice.Moreover, those two curves also coincide with the analyticalSER formulation corresponding to 𝑃 RD

𝑒,𝑘 , thus corroborating theclosed-form expression in (16).

Test Case 2 (S-MAP MUD algorithms). Next, the RD, LD,DDD, and SSD MUD algorithms are all tested for both BPSKand 4-PAM constellations, and their SER performance is com-pared. For LD, the quadratic program in (17) is solved usingthe general-purpose SeDuMi toolbox [13]. The quantizationrule chooses the nearest point in 𝒜𝑎 for both RD and LD. Forcomparison, we also include the ordinary LS detector, whichcorresponds to the RD solution in (10) with 𝜆 = 0.

Fig. 3(a) shows that the LS detector exhibits the worstperformance. This is intuitive since it neither exploits the finitealphabet nor the sparsity present in b. The SSD exhibits thebest performance at the price of highest complexity. The LDoutperforms the RD algorithm, as predicted. It is interestingto observe that even at low SNR region the DDD algorithmis surprisingly competitive, especially in view of its low com-plexity that grows only linearly in the number of symbols 𝐾 .The diversity orders for those detectors are basically the same.


5 10 15 20 2510

−4

10−3

10−2

10−1

Average SNR(dB)

SE

R

LSRDLDDDDSSD

(a)

5 10 15 20 25

10−3

10−2

10−1

Average SNR(dB)

SE

R

LSRDLDDDDSSD

(b)

Fig. 3. SER vs. average SNR (in dB) of sparsity-exploiting MUD algorithmswith 𝑁 = 32 and 𝐾 = 20 for (a) BPSK, and (b) 4-PAM alphabets.

This is reasonable since independent Rayleigh fading channelsbetween AP and users were simulated here. The correspondingcurves for 4-PAM depicted in Fig. 3(b) follow the same trend.However, compared to Fig. 3(a), the RD algorithm degradesnoticeably as its SER approaches the LS one. This is becausechoices of quantization thresholds become more influential asthe constellation size increases. As expected, the LD exhibitsresilience to this influence. The DDD and SSD algorithmshave almost identical performance in high-SNR region.

Test Case 3 ((In)activity across symbols). In this case, theuser (in)activity is correlated across time slots. We modelthis random (in)activity process as a two-state (active-inactive)stationary Markov chain. The state transition matrix is P =[𝑎 (1 − 𝑎); 𝑏 (1 − 𝑏)], where 𝑎 is uniformly distributed over[0.8 0.85], and 𝑏 over [0.05 0.1], for each user. For this model,the expected number of successive active slots is 1/(1 − 𝑎),and 1/𝑏 for the inactive ones. Also, the limiting probabilityfor the “active” state becomes 𝑏/(1 − 𝑎 + 𝑏), taking valuesfrom the interval [0.2 0.4]. Note that the activity factor overtime is still quite low. We use the RLS approach to estimate𝑝𝑎,𝑘(𝑡) as in (29) using 𝛽 = 0.5, and test both the DDD

0 5 10 15 20 25 30 35 40 45 50

10−3

10−2

Time slot t

SE

R

DDDSSD

SNR = 5 dB

SNR = 10 dB

SNR = 15 dB

Fig. 4. SER vs. time 𝑡 of sparsity-exploiting MUD algorithms with RLSestimation of the activity factors.

5 10 15 20 25

10−4

10−3

10−2

10−1

Average SNR(dB)

SE

R

RDDDDSSD

N = 32

N = 16

N = 8

Fig. 5. SER vs. average SNR (in dB) of sparsity-exploiting MUD algorithmswith 𝑁 = 32, 16, or 8 and 𝐾 = 20.

and SSD algorithms in solving the resultant S-MAP detectionproblem. The empirical SER is plotted in Fig. 4 across timefor different SNR values. Clearly, the proposed scheme iseffective in tracking the evolving time-correlated activity. Forthe same SNR value, it yields SER performance similar to theindependent case in Fig. 3(a).

Test Case 4 (Under-determined CDMA systems). We alsotest the S-MAP MUD algorithms for under-determined sys-tems, by varying 𝑁 from 32 to 16 and 8. The results aredepicted in Fig. 5. Since the RD is a simple linear detector,it is expected that once 𝑁 < 𝐾 , it will lose identifiability,and exhibits a considerably flat SER curve. At the same time,DDD still enjoys almost full diversity with a moderate choiceof 𝑁 = 16. Being the optimum detector, the SSD collects thefull diversity even if 𝑁 = 8; however, the other two kinds ofdetectors exhibit flat SER curves, as expected.

The Group Lasso scheme for recovering block activity isalso included for the under-determined case. Fig. 6 illustratesthe activity recovery error rate for different values of 𝑁 and𝑁𝑠. The number of observations 𝑁 affects the diversity order,


5 10 15 20

10−2

10−1

100

Average SNR(dB)

Err

or R

ate

N=16,N

s=20

N=16,Ns=10

N=16,Ns=1

N=8,Ns=20

N=8,Ns=10

N=8,Ns=1

Fig. 6. Error rate for block activity vs. average SNR (in dB) of Group Lassoalgorithm with 𝑁 = 16, 8 and 𝑁𝑠 = 20, 10, 1.

while the block size 𝑁𝑠 influences the recovery accuracy.

VIII. CONCLUSIONS

The MUD problem of sparse symbol vectors emerging withCDMA systems having low-activity factor was considered.Viewing user inactivity as augmenting the underlying alphabet,the a priori available sparsity information was exploited inthe optimal S-MAP detector. The exponential complexityassociated with the S-MAP detector was reduced by resortingto (sub-) optimal algorithms. Relaxed S-MAP detectors (RDand LD) come with low complexity but sacrifice optimality,because they ignore the finite-alphabet constraint. The secondkind of detectors (DDD and SSD), searches over a subset ofthe alphabet and exhibits improved performance at increasedcomplexity. The performance was analyzed for the RD andDDD algorithms, and closed-form expressions were derivedfor the corresponding symbol error rates.

S-MAP detectors were further generalized to deal withcorrelated user (in)activity across symbols by recursively esti-mating each user’s activity factor online; and also with under-determined (a.k.a. over-saturated CDMA) settings emergingwhen the spreading gain is smaller than the potential numberof users. Coping with the latter becomes possible throughregularization with the ℓ0 norm prior, or, with the GroupLasso-based recovery of the active user set. The numericaltests corroborated our analytical findings and quantified therelative performance of the novel sparsity-exploiting MUDalgorithms.

Our future research agenda includes analytical performanceevaluation in the under-determined case, which has beenavailable for the recovery of sparse signals but without finite-alphabet constraints. In addition, it will be interesting todevelop S-MAP detectors based on lattice search for higher-order constellations in the under-determined case.

APPENDIX APROOF OF (21)

To solve the optimization problem (20), consider first theunconstrained solution to the LS part of the cost, which can

be written as

𝑏LS𝑘 :=

(𝑦′𝑘 −

∑𝐾ℓ=𝑘+1 𝑅𝑘ℓ��

DDDℓ

)/𝑅𝑘𝑘. (32)

The detected symbol in (20) can be equivalently expressed as

��DDD𝑘 = arg min

𝑏𝑘∈𝒜𝑎

𝑓(𝑏𝑘),

𝑓(𝑏𝑘) :=(𝑏LS𝑘 − 𝑏𝑘

)2+ (2𝜆/𝑅2

𝑘𝑘)∣𝑏𝑘∣0. (33)

The solution ��DDD𝑘 can be obtained by comparing 𝑓(0) with

min𝑏𝑘∈𝒜 𝑓(𝑏𝑘). Specifically, as the cost 𝑓(𝑏𝑘) is quadratic for𝑏𝑘 ∈ 𝒜, the minimum is achieved at 𝑓(⌊𝑏LS

𝑘 ⌉), by quantizing𝑏LS𝑘 to the nearest point in 𝒜. Thus, ��DDD

𝑘 = 0 only if 𝑓(0) ≤𝑓(⌊𝑏LS

𝑘 ⌉), or equivalently, after using the definition of 𝑓(⋅), if2𝑏LS

𝑘 ⌊𝑏LS𝑘 ⌉ − ⌊𝑏LS

𝑘 ⌉2 − 2𝜆/𝑅2𝑘𝑘 ≤ 0. This completes the proof

of (21).

APPENDIX BPROOF OF (22)

When 𝑀 = 2, the DDD solution (21) reduces to

��DDD𝑘 = sign(𝑏LS

𝑘 )11(2∣𝑏LS𝑘 ∣ − 1 − 2𝜆/𝑅2

𝑘𝑘 > 0) (34)

due to the fact that ⌊𝑏LS𝑘 ⌉ = sign(𝑏LS

𝑘 ) ∈ {±1}.Recalling that b denotes the transmitted vector, and substi-

tuting y = Hb + w yields

y′ := Q𝑇y = Rb + u (35)

where u := Q𝑇w is zero-mean Gaussian with identity covari-ance matrix. Supposing that there is no error propagation, the𝑏LS𝑘 term in (34) becomes [cf. (32)]

𝑏LS𝑘 =

(𝐾∑ℓ=𝑘

𝑅𝑘ℓ��ℓ + 𝑢𝑘 −𝐾∑

ℓ=𝑘+1

𝑅𝑘ℓ��DDD𝑘

)/𝑅𝑘𝑘

= ��ℓ + 𝑢𝑘/𝑅𝑘𝑘 (36)

where 𝑢𝑘 denotes the 𝑘-th entry of u.To analyze the error probability for the detector in (34)

consider the following three cases.a) ��𝑘 = 0 is sent: under this case, 𝑏LS

𝑘 = 𝑢𝑘/𝑅𝑘𝑘 and adetection error emerges when ��DDD

𝑘 = 1 or −1. With theclosed-form DDD detector (34) in mind, such an error occursonly if both 𝑏LS

𝑘 ∕= 0 and 2∣𝑏LS𝑘 ∣ − 2𝜆/𝑅2

𝑘𝑘 − 1 > 0 hold.The first case corresponds to 𝑢𝑘 ∕= 0 and the second one isequivalent to ∣𝑢𝑘∣ > ∣𝑅𝑘𝑘∣/2 + 𝜆/∣𝑅𝑘𝑘∣, which is included inthe event 𝑢𝑘 ∕= 0. Hence, to evaluate the error probability itsuffices to consider only the case ∣𝑢𝑘∣ > ∣𝑅𝑘𝑘∣/2 + 𝜆/∣𝑅𝑘𝑘∣.b) ��𝑘 = 1 is sent: following the analysis in a), an error occursif sign(𝑅𝑘𝑘)𝑢𝑘 ≤ −∣𝑅𝑘𝑘∣/2 + 𝜆/∣𝑅𝑘𝑘∣.c) ��𝑘 = −1 is sent: following the analysis in a), an erroroccurs if sign(𝑅𝑘𝑘)𝑢𝑘 ≥ ∣𝑅𝑘𝑘∣/2 − 𝜆/∣𝑅𝑘𝑘∣.

Given that the Gaussian distributed 𝑢𝑘 has variance 1, theoverall SER for the 𝑘-th entry 𝑏𝑘 becomes

𝑃DDD𝑒,𝑘 =

∑𝑖=0,±1 𝑃 (error∣��𝑘 = 𝑖)𝑃 (��𝑘 = 𝑖)

= 2(1 − 𝑝𝑎)𝑄

( ∣𝑅𝑘𝑘∣2

+𝜆

∣𝑅𝑘𝑘∣)

+ 𝑝𝑎𝑄

( ∣𝑅𝑘𝑘∣2

− 𝜆

∣𝑅𝑘𝑘∣)

. (37)


ACKNOWLEDGEMENT

The authors wish to thank the anonymous reviewers for theirhelpful feedback which improved exposition of this paper.

REFERENCES

[1] D. Angelosante, E. Grossi, G. B. Giannakis, and M. Lops, “Sparsity-aware estimation of CDMA system parameters," EURASIP J. AdvancesSignal Process., June 2010, article ID 417981.

[2] R. G. Baraniuk, “Compressive sensing," IEEE Signal Process. Mag.,vol. 24, pp. 118-121, July 2007.

[3] C. R. Berger, S. Zhou, J. C. Preisig, and P. Willett, “Sparse channelestimation for multicarrier underwater acoustic communication: fromsubspace methods to compressed sensing," IEEE Trans. Signal Process.,vol. 58, no. 3, pp. 1708-1721, Mar. 2010.

[4] E. J. Candès and T. Tao, “Near-optimal signal recovery from randomprojections: universal encoding strategies?" IEEE Trans. Inf. Theory,vol. 52, pp. 5406-5425, Dec. 2006.

[5] E. Candès and M. B. Wakin, “An introduction to compressive sampling,"IEEE Signal Process. Mag., vol. 25, pp. 21-30, Mar. 2008.

[6] M. Cetin and B. M. Sadler, “Semi-blind sparse channel estimation withconstant modulus symbols," in Proc. Intl. Conf. Acoust., Speech, SignalProcess., Philadelphia, PA, Mar. 2005.

[7] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decompositionby basis pursuit," SIAM J. Scientific Comput., vol. 20, pp. 33-61, 1998.

[8] T. Cui and C. Tellambura, “An efficient generalized sphere decoder forrank-deficient MIMO systems," IEEE Commun. Lett., vol. 9, pp. 423-425, May 2005.

[9] M. O. Damen, H. E. Gamal, and G. Caire, “On maximum-likelihooddetection and the search for the closest lattice point," IEEE Trans. Inf.Theory, vol. 49, pp. 2389-2402, Oct. 2003.

[10] A. K. Fletcher, S. Rangan, and V. K. Goyal, “A sparsity detectionframework for on-off random access channels," in IEEE Proc. Intl.Symp. Inf. Theory, pp. 169-173, Seoul, South Korea, June 2009.

[11] G. B. Giannakis, Z. Liu, X. Ma, and S. Zhou, Space-Time Coding forBroadband Wireless Communications. John Wiley & Sons, Inc., 2007.

[12] A. H. Sayed, Fundamentals of Adaptive Filtering. John Wiley & Sons,2003.

[13] J. F. Sturm “Using SeDuMi 1.02, a MATLAB toolbox for optimizationover symmetric cones," Optimization Methods Software, vol.11-12, pp.625-653, 1999. [Online], Available: http://sedumi.mcmaster.ca.

[14] Z. Tian, G. Leus, and V. Lottici, “Detection of sparse signals underfinite-alphabet constraints," in Proc. Intl. Conf. Acoust., Speech SignalProcess., pp. 2349-2352, Taipei, Apr. 2009.

[15] R. Tibshirani, “Regression shrinkage and selection via the lasso," J.Royal Statistical Soceity: Series B, vol. 58, pp. 267-288, 1996.

[16] M. E. Tipping, “Sparse Bayesian learning and the relevance vectormachine," J. Machine Learning Research, vol. 1, pp. 211-244, 2001.

[17] S. Verdu, Multiuser Detection. Cambridge, UK: Cambridge Univ. Press,1998.

[18] E. Viterbo and J. Boutros, “A universal lattice code decoder for fadingchannels," IEEE Trans. Inf. Theory , vol. 45, pp. 1639-1642, July 1999.

[19] M. Yuan and Y. Lin, “Model selection and estimation in regression withgrouped variables," J. Royal Statistical Soceity: Series B, vol. 68, pp.49-67, 2006.

[20] H. Zhu and G. B. Giannakis, “Sparsity-embracing multiuser detectionfor CDMA systems with low activity factor," in Proc. IEEE Intl. Symp.Inf. Theory, pp. 164-168, Seoul, South Korea, June 2009.

Hao Zhu (S’07) received her Bachelor degree fromTsinghua University, Beijing, China in 2006 andthe M.Sc. degree from the University of Minnesota,Minneapolis in 2009, both in electrical engineering.

From 2005 to 2006, she was working as a researchassistant with the Intelligent Transportation informa-tion Systems (ITiS) Laboratory, Tsinghua Univer-sity. Since September 2006, she has been workingtowards her Ph.D. degree with the Department ofElectrical and Computer Engineering, University ofMinnesota. Her research interests include sparse

signal recovery and blind source separation in signal processing.

Georgios B. Giannakis (Fellow’97) received hisDiploma in Electrical Engr. from the Ntl. Tech. Univ.of Athens, Greece, 1981. From 1982 to 1986 he waswith the Univ. of Southern California (USC), wherehe received his MSc. in Electrical Engineering,1983, MSc. in Mathematics, 1986, and Ph.D. inElectrical Engr., 1986. Since 1999 he has been aprofessor with the Univ. of Minnesota, where he nowholds an ADC Chair in Wireless Telecommunica-tions in the ECE Department, and serves as directorof the Digital Technology Center.

His general interests span the areas of communications, networking andstatistical signal processing, subjects on which he has published more than300 journal papers, 500 conference papers, two edited books and two researchmonographs. Current research focuses on compressive sensing, cognitiveradios, network coding, cross-layer designs, mobile ad hoc networks, wirelesssensor and social networks. He is the (co-) inventor of 18 patents issued, andthe (co-) recipient of seven paper awards from the IEEE Signal Processing(SP) and Communications Societies, including the G. Marconi Prize PaperAward in Wireless Communications. He also received Technical AchievementAwards from the SP Society (2000), from EURASIP (2005), a Young FacultyTeaching Award, and the G. W. Taylor Award for Distinguished Research fromthe University of Minnesota. He is a Fellow of EURASIP, and has served theIEEE in a number of posts, including that of a Distinguished Lecturer for theIEEE-SP Society.

Documents

454 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 2